AI won’t replace you.
People who use it will.
The future is here. Adoption is near universal; depth is not, and that gap is the opportunity. Below is what four peer-reviewed experiments actually measured, including the findings that cut against the case for it, because you should not change how you work on somebody’s say-so.
AI Won’t Replace Humans, But Humans With AI Will Replace Humans Without AI
The question is not whether
it arrives.
It arrived. Stanford’s 2026 AI Index, drawing on McKinsey’s annual State of AI survey, puts organizational adoption at a level that ended the debate about timing.
Almost everyone has started. Almost nobody has finished. That gap is the whole opportunity, and it is the reason being early is still possible in 2026 even though the technology stopped being new.
The people who lose ground in the next two years will not lose it to a robot. They will lose it to the person in the next office who learned which parts of the job the tool is actually good at, and stopped doing those parts by hand.
Source: Stanford HAI, 2026 AI Index Report, Chapter 4, Economy, section 4.3, figure 4.3.1, sourced in the report to the McKinsey & Company Survey, 2025.
Four published experiments.
Same direction.
Not vendor case studies. Three are field experiments and one is a preregistered online experiment. Each is published in a peer-reviewed journal, each was randomized or staggered, and each states its sample size.
15% more issues resolved per hour
5,172 customer support agents, staggered rollout of an AI assistant. Productivity, measured as issues resolved per hour, rose 15% on average, “with substantial heterogeneity across workers.” That unevenness turns out to be the most useful thing in the paper, and it is covered two sections down.
Brynjolfsson, Li & Raymond, Generative AI at Work, Quarterly Journal of Economics, vol. 140 no. 2, May 2025.
40% less time, 18% better output
453 college-educated professionals, preregistered experiment, incentivized writing tasks from their own occupations. Average time taken fell 40%. Independently graded output quality rose 18%.
Noy & Zhang, MIT, Science, July 2023.
25.1% faster, 12.2% more done
758 knowledge workers at Boston Consulting Group, preregistered, randomly assigned to no AI, GPT-4, or GPT-4 plus a prompt engineering overview. Across 18 realistic consulting tasks inside the range of what the model does well, the AI groups completed 12.2% more tasks, completed them 25.1% more quickly, and delivered “solutions of significantly improved quality.” On one task deliberately chosen to sit outside that range, the same tool made them worse. See the next section but one.
Dell’Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon & Lakhani, Organization Science, vol. 37 no. 2, March to April 2026.
26% more tasks completed
Three field experiments pooled across 4,867 software developers using an AI coding assistant: a 26.08% increase in completed tasks. In the authors’ words, “less experienced developers had higher adoption rates and greater productivity gains.”
Cui, Demirer, Jaffe, Musolff, Peng & Salz, Management Science, 2026.
The edge is largest
for the person who is behind.
This is the finding that keeps repeating across separate studies in separate industries, and it is the opposite of what most people assume. AI does not reward the expert most. It rewards the person who has not got there yet.
One consistent finding across several of these studies is that the less experienced workers tended to benefit the most, suggesting that AI tools may help close existing skill gaps.
Both faster and better
What the least experienced support agents got, in the published paper’s exact words: “Less experienced and lower-skilled workers improve both the speed and quality of their output.” The earlier working paper version put the novice gain at 34% against a 14% average.
Faster, slightly worse
What the best agents in the same study got: “the most experienced and highest-skilled workers see small gains in speed and small declines in quality.” The tool is not a multiplier on mastery. The paper’s own explanation is that it spreads the best workers’ practices to everyone else, which helps you far more if you are not already one of them.
If you have been putting this off because you assume the people already ahead of you will benefit more, the research says you have it backwards. The gap it closes fastest is yours.
Where it goes wrong,
and it does go wrong.
A page that only shows the upside is an advertisement. Here is what the same body of research found when it looked for harm, because knowing this is the difference between using AI and being fooled by it.
19% less likely to be right
In the BCG experiment, one complex managerial task was chosen to sit outside what the model does well. Subjects using AI were “19% less likely to produce correct solutions” than those without it. The tool did not fail loudly. It failed persuasively. Worth knowing: the same paper states this result as “19 percentage points” in its discussion section and as “19%” in its abstract, so quote it carefully.
Experienced developers, 19% slower
METR found experienced open-source developers “became 19 percent slower when using AI assistance.” Stated with the caveat the AI Index attaches to it: METR could not replicate the result later, “primarily due to a growing reluctance among developers to work without AI, and that developers in late 2025 were likely sped up by AI relative to the original study period.” A real warning, not a settled fact, and a different study and measure from the 19% above.
Everyone starts, few finish
Adoption sits at 88% while scaled use of AI agents is in the single digits across nearly all business functions. Buying the subscription is not the same as changing how the work gets done, and most organizations have only done the first one.
The pattern across all three is one thing: the tool does not carry you. Knowing which tasks sit inside its range, and checking the ones near the edge, is the actual skill. That skill is learnable in an afternoon and almost nobody bothers.
Three steps, and you can
run them yourself this week.
Time the repetitive half first
Pick the task you do every week that you would never miss. Write down how long it takes and what a good result looks like. Do this before you touch a model, because without a before number you will never know whether anything improved, and you will believe it did.
Run it again with AI, and grade it
Same task, same definition of good. Measure the clock and the quality separately, never together. That is the whole lesson of the BCG study: on the task outside the model’s range, speed and correctness moved in opposite directions, and only a separate quality check would have caught it.
Keep what beat the clock. Drop the rest.
Some tasks will improve enormously and some will get quietly worse. Keep the winners, write down the losers so you stop retrying them, and move to the next task. That list of what works and what does not is the asset. The subscription is not.
Every number above,
and where it came from.
Named authors, named institutions, dated, and read in the version of record rather than in somebody’s summary of it. Check any of them.
Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond, “Generative AI at Work”
- Staggered introduction of a generative AI conversational assistant across 5,172 customer-support agents.
- Used above for: the 15% average gain in issues resolved per hour; “less experienced and lower-skilled workers improve both the speed and quality of their output”; and “the most experienced and highest-skilled workers see small gains in speed and small declines in quality.” All three are the published abstract’s own words.
- Version note, because the two circulate together. The 2023 NBER working paper reports 5,179 agents, a 14% average, and “minimal impact” on the most skilled. The peer-reviewed version supersedes it on all three, and the change from “minimal impact” to “small declines in quality” is a real finding, not a rewording, so this page uses the published numbers and cites the 34% novice figure to the working paper where it appears.
- academic.oup.com/qje/article/140/2/889/7990658 · working paper at nber.org/papers/w31161
Shakked Noy and Whitney Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence”
- Preregistered online experiment, 453 college-educated professionals, incentivized occupation-specific writing tasks, half randomly given ChatGPT.
- Used above for: “the average time taken decreased by 40% and output quality rose by 18%.”
- science.org/doi/10.1126/science.adh2586
Dell’Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon and Lakhani, “Navigating the Jagged Technological Frontier”
- Preregistered field experiment with 758 knowledge workers at Boston Consulting Group, randomly assigned to no AI, GPT-4, or GPT-4 with a prompt engineering overview.
- Used above for: 12.2% more tasks completed, 25.1% more quickly, and “solutions of significantly improved quality” across 18 tasks inside the frontier; and “19% less likely to produce correct solutions” on the one complex managerial task placed outside it. Every one of those is the abstract’s own wording.
- Two things worth knowing before you quote this paper elsewhere. First, the same result appears as “19%” in the abstract and as “19 percentage points” in the discussion, which are different quantities. Second, a widely repeated “more than 40% higher quality” figure appears nowhere in the published paper; its abstract says “significantly improved quality” and its discussion says “performance by more than 30%.” The 40% comes from the 2023 working paper abstract, which the published version dropped. This page therefore does not use it.
- pubsonline.informs.org/doi/10.1287/orsc.2025.21838 · open copy at hbs.edu/faculty/Pages/item.aspx?num=68729
Cui, Demirer, Jaffe, Musolff, Peng and Salz, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers”
- Three field experiments pooled, 4,867 developers.
- Used above for: “a 26.08% increase (SE: 10.3%) in completed tasks among developers using the AI tool” and “less experienced developers had higher adoption rates and greater productivity gains.”
- Note on a common misquote. The paper states the result in completed tasks. Several summaries render it as “26% more pull requests,” including the Stanford AI Index cited below. This page says completed tasks, which is what the abstract says. The quotation above is the 2025 working paper’s wording; the published version reads “standard error” where it reads “SE” and the number is unchanged.
- pubsonline.informs.org/doi/10.1287/mnsc.2025.00535 · working paper at papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566
Stanford HAI, 2026 AI Index Report, Chapter 4: Economy
- Used above for: 88% organizational adoption and 79% generative AI adoption, both sourced in the report to the McKinsey & Company Survey, 2025; scaled AI agent use in the single digits across nearly all functions; the METR finding that experienced open-source developers “became 19 percent slower when using AI assistance” together with the note that METR could not replicate it; and the verbatim line about less experienced workers benefiting most.
- One honesty note about this source. The report’s own chapter highlights state 70% for generative AI adoption while the body text and the data label on figure 4.3.1 both state 79%. This page uses 79%, the figure corroborated twice inside the document. Flagged rather than quietly picked.
- hai.stanford.edu/ai-index/2026-ai-index-report/economy
Karim R. Lakhani, “AI Won’t Replace Humans, But Humans With AI Will Replace Humans Without AI”
- Quoted above as the headline of that HBR piece, which is how it is published, rather than as a sentence recorded in his speech. Lakhani is a professor at Harvard Business School and a co-author of the BCG study cited above, so the line is not a slogan from outside the research. It is a researcher’s summary of his own. One disclosure, since this page holds others to the same standard: HBR sets that headline with a dash where this page uses a comma, the only change made to it.
- hbr.org/2023/08/ai-wont-replace-humans-but-humans-with-ai-will-replace-humans-without-ai